AI-generated responses, particularly those from large language models, can occasionally produce inaccurate or misleading information - a phenomenon known as “hallucination.”
At Mercu, we recognise this risk and have implemented a range of safeguards to ensure the accuracy, reliability, and fairness of AI outputs across our platform.
1. Clear Scope and Constrained Use Cases
We intentionally limit the role of AI to well-defined, tightly scoped functions. For example, AI is used to:
Check whether a candidate’s response is semantically relevant to a question
Generate a concise summary of a candidate’s transcribed video or voice interview
Assign a score (1–5) to an answer and explain the rationale
Answer candidate FAQs using provided job or company documentation
By clearly defining what the model is asked to do - and in what context - we reduce ambiguity, which is a common source of hallucinated outputs.
2. Retrieval-Augmented Generation (RAG) for FAQs
For our conversational candidate FAQ assistant, we don’t let AI guess. Instead, we use a Retrieval-Augmented Generation (RAG) pipeline, where the AI can only respond based on retrieved content from a predefined knowledge base - such as a job description, employee handbook, or company FAQ.
If the relevant answer isn’t found in the source material, the assistant is prompted to either ask for clarification or defer the question. This drastically limits hallucinations and keeps responses grounded in truth.
3. Human-in-the-Loop by Design
Mercu’s AI tools are designed to support human decision-making, not replace it. AI never makes final hiring decisions - such as whether to reject or progress a candidate. For every score or summary the system generates, we show the full underlying candidate response, allowing recruiters to review and override as needed. This ensures that the final decision remains in human hands, not driven solely by model output.
4. Data and Prompt Controls
We carefully craft prompts to be clear, deterministic, and tightly scoped to the task. Key prompts are customisable at the account level, allowing us to tune outputs based on customer needs while maintaining consistency and accuracy.
We also restrict what data the model can see. For example, scoring and summaries are generated purely from text transcripts - without access to audio tone, video appearance, or personally identifying information like name, age, gender, or race (unless voluntarily disclosed by the candidate). This not only reduces bias but also reduces the risk of the AI making speculative inferences.
5. Ongoing Monitoring and Testing
All AI interactions are logged and monitored internally for traceability and quality assurance. We continuously test AI responses - especially in production environments - to catch and correct any drift in behaviour. If hallucinations or unexpected outputs are detected, we revise the prompt, update the retrieval context, or temporarily disable the feature if needed.
By combining prompt discipline, limited autonomy, retrieval-based methods, and human oversight, we ensure that our AI features are accurate, explainable, and safe.